Wiki Pages

Agent Evaluation & Benchmarks

Agent Memory

GUI Agents

Multi-Agent Systems

Tool Use & Function Calling

AI for Auto-Research: Roadmap & User Guide

Look Before You Leap: Autonomous Exploration for LLM Agents

MMSkills: Multimodal Skills for General Visual Agents

PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control

Solvita: Agentic Evolution for Competitive Programming

LIFE Survey: Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems

FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale

SPIN: Structural LLM Planning via Iterative Navigation for Industrial Tasks

Agent Memory Cluster: STALE + Preping + EvolveMem + MemEye + MemLens + BOOKMARKS

EvoEnv: Self-Evolving Reasoning RL via Verifiable Environment Synthesis

Orchard: Open-Source Agentic Modeling Framework — 67.5% SWE-bench Verified at 30B

SDAR: Self-Distilled Agentic Reinforcement Learning

WildClawBench: Native-Runtime Long-Horizon Agent Benchmark — Claude Opus 4.7 Tops Out at 62.2%

AgentLens: the Lucky Pass problem in SWE-agent evaluation

Context Training with Active Information Seeking

Revisiting DAgger in the era of LLM agents

MAP: a Map-then-Act paradigm for long-horizon interactive agents

Agent-BRACE: Decoupling Beliefs from Actions in Long-Horizon Tasks

LLM Agents Already Know When to Call Tools — Even Without Reasoning (Probe&Prefill)

LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues

Useful Memories Become Faulty When Continuously Updated by LLMs

X-OmniClaw: Unified Mobile Agent for Multimodal Understanding and Interaction

AutoTTS: Agentic Discovery for Test-Time Scaling

Jiayi Weng: Learning Beyond Gradients

OncoAgent: Dual-tier multi-agent framework for privacy-preserving oncology decision support

AI Co-Mathematician

Auto Research with Specialist Agents Develops Effective and Non-Trivial Training Recipes

Beyond Semantic Similarity: Direct Corpus Interaction (DCI)

Skill Curation Cluster: StraTA, Skill1, SkillOS

BRIGHT-Pro and RTriever-4B: Reasoning-Intensive Retrieval for Agentic Search

MedSkillAudit: Domain-Specific Audit Framework for Medical Research Agent Skills

OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents

AcademiClaw: When Students Set Challenges for AI Agents

Ctx2Skill: From Context to Skills — Self-Evolving Multi-Agent Skill Extraction

PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments

T^2PO: Token- and Turn-Level Policy Optimization for Stable Multi-Turn Agentic RL

Why Your Agentic AI Pentester Is Probably Just a Fancy Scanner — Ken Huang

LWD — Learning While Deploying: Fleet-Scale RL for Generalist Robot Policies

Ken Huang Ch 15 — Structured Output and Schema-Constrained Generation

Ara: Agent-Native Research Artifacts

Claw-Eval-Live: Live Agent Benchmark for Evolving Real-World Workflows

Eywa: Heterogeneous Scientific Foundation Model Collaboration

InteractWeb-Bench: Multimodal Agents under Non-Expert User Instructions

Intern-Atlas: A Methodological Evolution Graph

MCP Integration: Claude Code vs Hermes (Ken Huang Ch 13)

Synthetic Computers at Scale: Long-Horizon Productivity Simulation

ClawGym: A Scalable Framework for Building Effective Claw Agents

Claude Code Memory Systems: Chapter 8 Analysis

Claude Code vs. Hermes Agent: Permission System Architectures

Persistent Agent Infrastructure: Kimi K2.6, OpenAI Agent Studio, Anthropic Conway

AgentSPEX: An Agent Specification and Execution Language

SimpleTES: Evaluation-Driven Scaling for Scientific Discovery

HuggingFace ml-intern: Open-Source Agentic Post-Training Loop

Precise Debugging Benchmark: Models Regenerate, They Don't Debug

Reward-Free Self-Evolution: Agents That Learn Without Being Told What to Learn

GTA-2: Benchmarking General Tool Agents from Atomic Use to Open-Ended Workflows

PRL-Bench: LLMs on Frontier Physics Research

Chapter 3: The Query/Agent Loop — Claude Code vs. Hermes Agent

Claude Code Architecture: A Deep Reading

UniDoc-RL: RL-Based Visual RAG with Hierarchical Actions

Corpus2Skill: Don't Retrieve, Navigate

DR3-Eval: Realistic Benchmark for Deep Research Agents

Dive into Claude Code: Architecture Analysis

SuperLocalMemory V3.3: Biologically-Inspired Agent Memory

DefenseClaw, MAESTRO, and the Security Boundary Agentic AI Has Been Missing

Do AI Coding Agents Log Like Humans? An Empirical Study

Exploration and Exploitation Errors Are Measurable for Language Model Agents

OccuBench: Evaluating AI Agents on Real-World Professional Tasks via Language World Models

TREX: Automating LLM Fine-tuning via Agent-Driven Tree-based Exploration

UI-Copilot: Long-Horizon GUI Automation via Tool-Integrated Policy Optimization

Inside VAKRA: Reasoning, Tool Use, and Failure Modes of Agents

Open Artifacts #21: The May 2026 Open-Model Wave and the CAISI / ECI Gap

Anthropic overtakes OpenAI in B2B adoption for the first time (Ramp data)

Recursive emerges from stealth with $650M for self-improving AI

Broadcom won't build OpenAI's custom chip without Microsoft buying 40 percent

Anthropic ↔ Colossus 1 Deal: Capacity Crunch + Brand Risk

GitHub Reliability Crisis: AI Load Breaks the Platform

Lambert: Notes from inside China's AI labs

Anthropic + OpenAI Both Build Services Companies Around Their AI

Microsoft VS Code Auto-Inserts "Co-Authored-by Copilot" Even With AI Off

Xiaomi MiMo-V2.5-Pro — Open-Weight Long-Horizon Coding at 40-60% Fewer Tokens

ChatGPT Tracks Free Users for Ads by Default

AISN #72 — CAIS AI-Wellbeing Research, Public-Sentiment Decline, OpenAI Releases

Anthropic Launches Claude Security — Defensive Cyber Productization

Chinese AI Startups Onshoring: Moonshot, StepFun Dissolving Offshore Structures

UK AISI: GPT-5.5 Matches Claude Mythos on Full Network Attack Simulation

Marcus: "The Greatest Capital Misallocation in History?"

Pentagon Signs Eight Tech Giants for AI-First Fighting Force; Anthropic Excluded

Pragmatic Engineer: AI Load Breaks GitHub; Anthropic's Trust Speedrun

SemiAnalysis: AI Value Capture — The Shift to Model Labs

Zig's Anti-LLM Contribution Policy and the "Contributor Poker" Argument

Claude for Creative Work: MCP Connectors for Blender, Adobe, Ableton, Autodesk

UCP Wins the Agentic Commerce Governance Layer

Amazon-Anthropic $33B Deal and AI Capital Concentration Week

OpenAI Executive Departures and Product Restructuring (April 2026)

Anthropic's Mythos Model: Government Access and the Trust Debate

Claude's Explosive Market Share Surge (April 2026)

LLM Routing

MoE-muP: How to Scale Mixture-of-Experts (From muP to the Maximally Scale-Stable Parameterization)

BEAM: Binary Expert Activation Masking for Dynamic Routing in MoE

Dynamic Latent Routing: Joint Latent-Code and Routing Policy for LM Post-Training

RouteProfile: Elucidating the Design Space of LLM Profiles for Routing

CaRE: Scaling Continual Learning to 300+ Tasks with Bi-Level Routing Mixture-of-Experts

Conductor: Learning to Orchestrate Agents in Natural Language

State of Routing in Model Serving (Netflix Tech Blog)

Step-level Optimization for Efficient Computer-use Agents

Ken Huang Ch 14: Model Routing and Provider Abstraction (Claude Code vs Hermes)

TRACER: Trace-Based Adaptive Cost-Efficient Routing for LLM Classification

GPU Kernels and Accelerator Optimization

AgentKernelArena: Generalization-Aware Benchmarking of GPU Kernel Optimization Agents

Inference as energy-to-token production: a position paper

SemiAnalysis: Cerebras — Faster Tokens Please

KernelBench-X: A Comprehensive Benchmark for Evaluating LLM-Generated GPU Kernels

Cerebras Targets $40B Valuation in Second IPO Attempt

Semiconductor Week 17, 2026: AI Memory Supercycle and Agentic EDA

SemiAnalysis: GPU Cluster Economics and the Goodput Reckoning

Knowledge Distillation

KV Cache

Speculative Decoding

CompactAttention: Accelerating Chunked Prefill with Block-Union KV Selection

EndPrompt: Efficient Long-Context Extension via Terminal Anchoring

LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation

Measuring Maximum Activations in Open Large Language Models

PUMA: Semantic-Preserving Early Exit for Reasoning Models

SNLP: Layer-Parallel Inference via Structured Newton Corrections

ZEDA: Post-Trained MoE Can Skip Half Experts via Self-Distillation

FashionChameleon: Training-Free KV Cache Rescheduling for Interactive Video Customization

HodgeCover: Higher-Order Topological Coverage Drives Compression of Sparse Mixture-of-Experts

MTP support merged into llama.cpp: Strix Halo benchmarks confirm a 2x decode speedup at 27B, mixed result at 35B

Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention (Raschka)

ATESD: Adaptive Teacher Exposure for Self-Distillation in LLM Reasoning

Lighthouse Attention: Long-Context Pre-Training as a Detachable Wrapper

Asynchronous Continuous Batching: CPU-GPU Overlap via Dual Buffer Slots

Forcing-KV: Hybrid KV Cache Compression for Autoregressive Video Diffusion

The Extrapolation Cliff: a closed-form clip-safety threshold for on-policy distillation

MinT: managed infrastructure for million-scale LoRA training and serving

MMProLong: training long-context vision-language models with generalization beyond 128K

Orthrus: dual-view diffusion + autoregressive on a shared KV cache

δ-mem: Efficient Online Memory for Large Language Models

FocuSFT: Bilevel Optimization for Dilution-Aware Long-Context Fine-Tuning

The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes

Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle

Token Superposition Training (TST): Efficient Pre-Training with Token Superposition

Make Each Token Count: Improving Long-Context Performance with Learned KV Eviction

MDN: Parallelizing Stepwise Momentum for Delta Linear Attention

MISA: Mixture of Indexer Sparse Attention for Long-Context LLM Inference

UniPrefill: Universal Long-Context Prefill Acceleration via Block-wise Dynamic Sparsification

EMO: Pretraining Mixture of Experts for Emergent Modularity

MiA-Signature: Approximating Global Activation for Long-Context Understanding

UniPool: A Globally Shared Expert Pool for Mixture-of-Experts

D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models

LIVEditor: Lightning Unified Video Editing via In-Context Sparse Attention (ISA)

Stream-R1: Reliability-Perplexity Aware Reward Distillation for Streaming Video Generation

Stream-T1: Test-Time Scaling for Streaming Video Generation

MotionCache: Motion-Aware Caching for Efficient Autoregressive Video Generation

The Distillation Panic — Nathan Lambert (Interconnects AI)

FlashRT: Efficient Red-Teaming for Prompt Injection and Knowledge Corruption

Nemotron 3 Nano Omni: Efficient Open Multimodal Intelligence

LenVM: Token-Level Length Value Model

RoundPipe: Efficient Training on Multiple Consumer GPUs

Accelerating RL Post-Training Rollouts via System-Integrated Speculative Decoding

Tide: Cross-Architecture Distillation for Diffusion Large Language Models

PrfaaS: Prefill-as-a-Service via Cross-Datacenter KV Cache Transfer

SDVG: Speculative Decoding for Autoregressive Video Generation

ShadowPEFT: Centralized Layer-Space Parameter-Efficient Fine-Tuning

TurboQuant: Online Vector Quantization for KV Cache Compression

Nemotron 3 Super: Hybrid Mamba-Attention MoE at NVFP4

1D Ordered Tokens Enable Efficient Test-Time Search

AccelOpt: Self-Improving LLM Agent for AI Accelerator Kernel Optimization

AVR: Adaptive Visual Reasoning for Efficient VRMs

Maximal Brain Damage: Disrupting Neural Networks via Sign-Bit Flips

STOP: Super Token for Path Pruning in Parallel Reasoning

W-RAC: Web Retrieval-Aware Chunking for Cost-Efficient RAG

LongAct: Harnessing Intrinsic Activation Patterns for Long-Context RL

Switch-KD: Visual-Switch Knowledge Distillation for VLMs

TESSY: Teacher-Student Cooperation Framework for SFT Data Synthesis

Cross-Tokenizer LLM Distillation via Byte-Level Interface

KV Packet: Recomputation-Free Context-Independent KV Caching

Model Capability Dominates: Lessons from AIMO 3 Inference-Time Optimization

TIP: Token Importance in On-Policy Distillation

Reinforcement Learning for LLMs

DiHAL: Where Should Diffusion Enter a Language Model? Geometry-Guided Hidden-State Replacement

MixSD: Mixed Contextual Self-Distillation for Knowledge Injection

NGM: A Plug-and-Play Training-Free Memory Module for LLMs

AIRA-Compose and AIRA-Design: Agentic Discovery of Neural Architectures

CIPO: Correction-Oriented Policy Optimization with Verifiable Rewards

NudgeRL: Strategy-Guided Exploration for RLVR

Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Reasoning

SU-01: Gold-Medal Olympiad Reasoning at 30B via Simple and Unified Scaling

Many-Shot CoT-ICL: long context as structured curriculum, not retrieval buffer

Reward Hacking in Rubric-Based Reinforcement Learning

G-Zero: Self-Play for Open-Ended Generation from Zero Data

Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training

Model Merging Scaling Laws in Large Language Models

Rebellious Student: Reversing Teacher Signals for Reasoning Exploration with Self-Distilled RLVR

Soohak: Mathematician-Curated Research-Level Math Benchmark

Gowers + ChatGPT 5.5 Pro: PhD-level math research in under two hours

Balanced Aggregation: Understanding and Fixing Aggregation Bias in GRPO

Prescriptive Scaling Laws for Data Constrained Training

TIDE: Every Layer Knows the Token Beneath the Context

ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning

When to Think, When to Speak: Learning Disclosure Policies for LLM Reasoning

Import AI 455: AI Systems Are About to Start Building Themselves — Jack Clark

Themis — Robust Multilingual Code Reward Models for Multi-Criteria Scoring

Ken Huang — World Models, Architectures, and the Next Phase of AI

Marcus — Have LLMs Improved Patient Outcomes?

MIT Study — Superposition Explains Why Scaling Language Models Works So Reliably

Philosophy-Bench — Frontier Models Diverge on 100 Everyday Ethical Scenarios

ARC-AGI-3 — Three Systematic Reasoning Errors in Frontier Models

Compliance vs Sensibility: Reasoning Controllability in LLMs

The Defense Trilemma + NP-Hardness of Reward Hacking Detection

Safety Drift After Fine-Tuning: Evidence from High-Stakes Domains

CoPD: Co-Evolving Policy Distillation

A Survey on LLM-Based Conversational User Simulation

Hope Architecture: Nested Learning and Continuously Adapting LLMs

DeepSeek V4: Architecture and Industry Impact

GPT-5.5: Launch Analysis and System Card Deep Dive

Chain-of-Thought Degrades Visual Spatial Reasoning

Target-Oriented Pretraining via Neuron-Activated Graph (NAG)

TEMPO: Scaling Test-Time Training for Large Reasoning Models

Weight Disentanglement and Task Arithmetic: OrthoReg

Geometric Canary: Steerability and Drift Detection from Representational Geometry

GFT: SFT is Degenerate Policy Gradient — and Group Fine-Tuning Fixes It

When Does RLVR Generalize? Reward Saturation and Reasoning Faithfulness

ASGuard: Mechanistic Defense Against Targeted Jailbreaking

Value Gradient Flow: RL as Optimal Transport

C2: Cooperative-Critical Rubric-Augmented Reward Modeling

InfiniteScienceGym: Procedurally-Generated Benchmark for Scientific Analysis

My Bets on Open Models, Mid-2026

From P(y|x) to P(y): Reinforcement Learning in Pre-train Space

Responsible AI

Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics

DiagnosticIQ: LLM Deployment Calibration Bottleneck in Industrial Maintenance

LLM-based Detection of Manipulative Political Narratives

LiSA: Lifelong Safety Adaptation via Conservative Policy Induction

WriteSAE: sparse autoencoders for the recurrent matrix cache write

A Single Layer to Explain Them All: Understanding Massive Activations in LLMs (ME Layer)

Pseudoscientific emotion AI in the workplace (Atlantic via The Decoder)

Anthropic Natural Language Autoencoders (NLAs)

The First Token Knows: Single-Decode Confidence for Hallucination Detection

2026-05-19-afternoon

2026-05-19-evening

2026-05-19-morning

2026-05-19

2026-05-18-afternoon

2026-05-18-evening

Media Live | 2026-05-18 morning slot

2026-05-18

2026-05-17-afternoon

2026-05-17-evening

Media Live | 2026-05-17 morning

2026-05-17

2026-05-16-afternoon

2026-05-16-evening

2026-05-16-morning

2026-05-16

2026-05-15-afternoon

2026-05-15-evening

2026-05-15-morning

2026-05-15

2026-05-14-afternoon

2026-05-14-evening

2026-05-14-morning

2026-05-14

2026-05-13-afternoon

2026-05-13-evening

2026-05-13

2026-05-12-afternoon

2026-05-12-evening

2026-05-12-morning

2026-05-12

2026-05-11-afternoon

2026-05-11-evening

2026-05-11-morning

2026-05-11

2026-05-10-afternoon

2026-05-10-evening

2026-05-10-morning

2026-05-10

2026-05-09-morning

2026-05-09

2026-05-08-night

2026-05-08

2026-05-07-afternoon

2026-05-06-morning

2026-05-06-night

CurveBench: Hierarchical Topological Reasoning from Visual Input

Auto-Rubric as Reward (ARR): From Implicit Preferences to Explicit Multimodal Generative Criteria

DeltaRubric: Generative Multimodal Reward Modeling via Joint Planning and Verification

ROMA: Reinforcing Multimodal Reasoning Against Visual Degradation

APEX: Aesthetic-Informed Popularity Prediction for AI-Generated Music

HERMES++: Unified Driving World Model for 3D Scene Understanding and Generation

JoyAI-Image: Awaking Spatial Intelligence in Unified Multimodal Understanding and Generation

Parameter-Efficient Multi-View Proficiency Estimation

PhysForge: Physics-Grounded 3D Asset Generation

RLDX-1: VLA Robotic Policy for Dexterous Humanoid Manipulation

AnalogRetriever — Cross-Modal Representations for Analog Circuit Retrieval

End-to-End Autoregressive Image Generation with 1D Semantic Tokenizer

GenLIP — Generative Language-Image Pre-training for ViTs

Map2World — Segment Map Conditioned Text-to-3D World Generation

UniVidX — Unified Multimodal Framework for Versatile Video Generation

Nemotron 3 Nano Omni — Efficient Open Multimodal Intelligence (NVIDIA)

Semi-DPO: Learning from Noisy Preferences via Semi-Supervised DPO

ViPO: Visual Preference Optimization at Scale

Edit-R1: Verifier-Based RL for Image Editing

FD-loss: Representation Fréchet Loss for Visual Generation

PhyCo: Controllable Physical Priors for Generative Motion

Visual Generation in the New Era: Atomic to Agentic World Modeling

Diffusion Templates: A Unified Plugin Framework for Controllable Diffusion

FASH-iCNN: Editorial Fashion Identity via Multimodal CNN Probing

GLM-5V-Turbo: Native Foundation Model for Multimodal Agents

X-WAM: Unified 4D World Action Modeling with Asynchronous Denoising

Qwen3.5-Omni Technical Report

GameWorld: Standardized and Verifiable Evaluation of Multimodal Game Agents

MERRIN: Multimodal Evidence Retrieval and Reasoning in Noisy Web Environments

RationalRewards: Reasoning Rewards Scale Visual Generation at Training and Test Time

Seedance 2.0: Advancing Video Generation for World Complexity